To help cut time to market, optimize production and extend component life, MINES ParisTech’s industry partners wanted a breakthrough solution combining computational fluid dynamics with deep learning.
MINES ParisTech deployed a hybrid platform based on an IBM Power Systems AC922 server with two IBM POWER9 processors and two NVIDIA V100 GPUs connected with NVIDIA NVLink technology.
2X boostin computational performance out of the box
CPU-to-GPU coherenceenables exceptional hybrid performance
Speedscutting-edge industrial research to help partners with digital transformation
Business challenge story
As the leading academic institute for collaborative industry research in France, MINES ParisTech maintains close, mutually beneficial links with dozens of major companies. The goal is to ensure that research contributes to advancements in industry, which then feed back into new research requirements. This creates a virtuous circle that provides world-class opportunities for students and researchers as well as significant competitive advantage for participating companies.
Professor Elie Hachem, Head of the Computing and Fluids Research Group at CEMEF MINES ParisTech, says, “As part of the ongoing digital transformation of industries, companies are seeking to optimize the production cycle, reduce time to market, transform manufacturing and extend the service life of their products. To help our partners achieve these goals, we recognized that we needed a new technology platform that could support the convergence of computational fluid dynamics [CFD] and machine learning.”
As Hachem observes, this convergence is a logical next step: classic computer simulations generate enormous volumes of data that represent the ideal raw material for the latest deep-learning algorithms. In this way, rather than simply serving to answer the original research challenge, big data from simulations can be reused to unlock new insights.
Through its research into the convergence of CFD and machine learning, MINES ParisTech has three key objectives. First, to augment its CFD capabilities with artificial intelligence (AI) in order to increase the fidelity of simulations and minimize the existing gap between simulations and physical reality. Second, to dramatically accelerate predictive modeling by reusing data from classical simulations in deep-learning algorithms. And third, to accelerate the optimization of industrial and manufacturing controls through hybrid computing.
“To truly converge CFD and machine learning, we needed a seamless hybrid platform offering the same high performance for both types of computation,” says Hachem. “We also wanted a packaged solution that would be easy for our industry partners to deploy and support.”
Working with its long-term technology partner Carri Systems, MINES ParisTech deployed an IBM Power Systems AC922 server with two IBM POWER9 processors and two NVIDIA V100 GPUs connected with NVIDIA NVLink technology. NVLink is a direct GPU-to-GPU interconnect technology that offers ten times the bandwidth of PCIe Gen 3 for ultra-fast deep-learning training.
“We chose the IBM Power AC922 server for three key reasons: its small footprint, its hybrid CPU/GPU performance and its scalability,” says Aurélien Larcher, Assistant Professor in Hachem’s research group. “The new POWER9 processors doubled our computational performance even before we conducted any code optimization. And the links with the GPU work really well – it’s a very well-balanced hybrid solution. As ever, Carri Systems did an excellent job to get us up and running in less than one week.”
Designed to support AI, the IBM Power AC922 features CPU-to-GPU coherence that allows system memory to be directly addressed as GPU memory, accelerating hybrid computations by reducing data movement and locality requirements. The server is the backbone of the US Department of Energy's Summit and Sierra supercomputers, which topped the most recent (November 2019) TOP500 list of the world’s fastest computing platforms.
MINES ParisTech is using its Power AC922 server to support student research and the development of new hybrid CPU/GPU software solutions for industry within the MINDS project (MINES Initiative for Numerics and Data Science), supported by CARNOT MINES. This project brings together 15 research centers across MINES ParisTech and six other institutions, with the goal of providing an AI-as-a-service platform to industry partners.
“The work we are doing on hybrid architectures with the Power AC922 will help us to improve simulation fidelity while cutting processing time, and to democratize this type of service for use by companies of all types,” says Hachem. “This was another factor in our selection. We could have created our own custom hybrid server, but that would have imposed a heavy burden on our industry partners, because they would then have needed to deploy and support a non-standard machine. The IBM Power AC922 is a compact, pre-integrated server that is powerful but also manageable, standardized and supported by a major global company.”
With more than two times the performance on the CPU side plus tightly integrated GPU capabilities, the IBM Power AC922 gives MINES ParisTech an exceptionally capable machine for both CFD and machine learning. In addition to cutting the computation time for tasks such as modeling the airflow over jet engine components, the IBM solution will help MINES ParisTech bring new predictive analytics tools to its industrial partners.
“Currently, it takes perhaps four years to bring a new car to market and seven years for an aircraft,” says Hachem. “Our partners want to accelerate processes throughout the product cycle from concept to production, and the new approach will be to combine traditional computer simulation with real-time machine learning. With the Power AC922, we are working with clients across transportation, energy, metallurgy, oil and gas, manufacturing and more to speed innovation and reduce time to market.”
By combining CPU and GPU computations in a single physical system, MINES ParisTech expects to enable faster and more accurate modeling as well as optimization through predictive analytics. Ultimately, the solution will help the institute’s industry partners to increase their operating efficiency and cut time to market.
“An example use case for the hybrid architecture is modeling heat flow in an aircraft component,” says Hachem. “To ensure safe operation and maximize service life, it’s important to minimize temperature gradients and eliminate hot spots. By combining classic CFD on the CPU with deep reinforcement learning [DRL] on the GPU, we can rapidly zero in on the optimal design parameters without having to model a huge number of scenarios.”
The CFL group at CEMEF – MINES ParisTech has already set a new benchmark on thermal transfer control by forced convection. The objective is to cool down a heated object using three inlet nozzles of fixed cross-section, located at the upper boundary of the domain. The CFD-DRL coupling makes it possible to dynamically test various positions for the jet inlets while aiming to minimize the temperature gradients in order to ensure uniform cooling. Figure 1 (above) shows snapshots of conjugate heat transfer and fluid flow for various positions of the three jet inlets. Interestingly, it can be seen that convergence is achieved once the DRL superposes two of the three inlets and pushes them against the lateral boundaries, as shown in the last sub-figure, which results in perfect symmetric cooling of the centered component.
“With the IBM Power AC922, the coupling between CPU and GPU is effective and transparent, so you don’t have to finish a computation on one and then wait for the other to finish,” says Hachem. “As our industry partners tackle challenges around the digital transformation of their factories and processes, we’re ready to support them with this hybrid IBM Power architecture that will deliver significant improvements in performance and new predictive capabilities.”
Established over 230 years ago, MINES ParisTech (external link) is a leading French graduate school that houses 18 research centers in five fields: earth sciences and environment, energy and process engineering, mechanical and materials engineering, mathematics and systems, and economics, management and society. It has more than 1,280 students and locations in Paris, Évry, Fontainebleau and Sophia Antipolis.